joycraft 1.3.0 → 1.5.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (33) hide show
  1. package/README.md +154 -4
  2. package/dist/joycraft-release.json +1 -1
  3. package/package.json +3 -3
  4. package/payload/.agents/skills/joycraft-context/SKILL.md +5 -6
  5. package/payload/.agents/skills/joycraft-decompose/SKILL.md +9 -8
  6. package/payload/.agents/skills/joycraft-entry/SKILL.md +8 -19
  7. package/payload/.agents/skills/joycraft-implement/SKILL.md +16 -17
  8. package/payload/.agents/skills/joycraft-session-end/SKILL.md +10 -15
  9. package/payload/.agents/skills/joycraft-testing/SKILL.md +3 -9
  10. package/payload/.agents/skills/joycraft-verify/SKILL.md +20 -21
  11. package/payload/.claude/agents/joycraft-implementer.md +5 -5
  12. package/payload/.claude/agents/joycraft-reviewer.md +3 -6
  13. package/payload/.joycraft/contracts/types.d.ts +18 -0
  14. package/payload/.joycraft/references/architecture.md +2 -4
  15. package/payload/.joycraft/references/completion.md +16 -21
  16. package/payload/.joycraft/references/context.md +11 -34
  17. package/payload/.joycraft/references/dispatch.md +7 -40
  18. package/payload/.joycraft/references/gates.md +60 -0
  19. package/payload/.joycraft/references/retro.md +6 -15
  20. package/payload/.joycraft/references/settings.md +6 -3
  21. package/payload/.joycraft/references/setup.md +4 -4
  22. package/payload/.joycraft/references/testing.md +21 -21
  23. package/payload/.joycraft/references/worktrees.md +1 -1
  24. package/payload/.joycraft/runtime/completion.mjs +92 -6
  25. package/payload/.joycraft/runtime/dispatch-preflight.mjs +1 -1
  26. package/payload/.joycraft/runtime/dispatch-start.mjs +61 -4
  27. package/payload/.joycraft/runtime/evidence.mjs +134 -0
  28. package/payload/.joycraft/runtime/gate.mjs +308 -0
  29. package/payload/.joycraft/runtime/identity.mjs +5 -1
  30. package/payload/.joycraft/runtime/queue-check.mjs +355 -0
  31. package/payload/.joycraft/runtime/status.mjs +46 -5
  32. package/payload/PIN +31 -28
  33. package/payload/scripts/export.mjs +5 -2
@@ -65,6 +65,8 @@ export interface QueueSpec {
65
65
  /** Captured before implementation; never infer it from a moving upstream branch. */
66
66
  changeScope?: { baseRevision: string; exclusions?: Array<{ path: string; reason: string }> };
67
67
  };
68
+ /** Written by `completion.mjs submit` with the move to in-review; subject bindings leave it out, like status. */
69
+ submitted?: { at: string };
68
70
  acceptedReport?: { path: string; sha256: string };
69
71
  /** Explicit owner exception, mutually exclusive with acceptedReport; not a PASS. */
70
72
  ownerAcceptance?: Receipt;
@@ -144,8 +146,24 @@ export interface VerificationReport {
144
146
  verdict: Result;
145
147
  flows: FlowEvidence[];
146
148
  unresolved: string[];
149
+ /** Optional; a report without it reads as before. */
150
+ findings?: ReviewFinding[];
147
151
  createdAt: string;
148
152
  }
153
+ /** A follow-up never blocks; a blocking or decision-needed finding refuses completion (exit 2). */
154
+ export interface ReviewFinding {
155
+ /** Unique in the report. */
156
+ id: string;
157
+ class: 'blocking' | 'decision-needed' | 'follow-up';
158
+ origin: 'new' | 'pre-existing' | 'unknown';
159
+ /** A criterion id of the spec. A blocking or decision-needed finding needs `criterion` or `at`. */
160
+ criterion?: string;
161
+ /** file:line or a receipt path. */
162
+ at?: string;
163
+ summary: string;
164
+ /** The reading the reviewer took for a case the spec does not decide. */
165
+ assumption?: string;
166
+ }
149
167
  /** Scoped human instruction recorded by the host; bindings do not authenticate its origin. */
150
168
  export interface OwnerAcceptance {
151
169
  schemaVersion: 1;
@@ -79,10 +79,8 @@ subject or provide the reviewer with an intended verdict or implementer reasonin
79
79
  Update only affected accepted facts in their owning reference, reconciling them
80
80
  at session end before any eligible archival. Active choices remain in their
81
81
  brief; rationale links to that choice; proposals stay marked as proposals.
82
- References describe and navigate the project. They neither grant permission nor
83
- become global Jev constraints. Do not add them to global hook rule paths or make
84
- extra live classification calls merely to maintain a map. Retrieval scores and
85
- this procedure do not certify exhaustive discovery.
82
+ References describe and navigate the project. They do not grant permission.
83
+ Retrieval scores and this procedure do not certify exhaustive discovery.
86
84
 
87
85
  Keep task context bounded to relevant sources. When evaluating the change,
88
86
  record newly delivered input bytes, repeated reads/provider calls, human
@@ -1,12 +1,13 @@
1
1
  # Guarded completion
2
2
 
3
- Use this reference when snapshotting a verification subject, submitting an independent review report, or responding to a completion failure.
3
+ Use this reference to submit a spec, snapshot a subject, finalize a review or recover a write.
4
4
 
5
5
  ## Stable local interface
6
6
 
7
7
  Run these commands from the consumer project root. Paths are project-relative.
8
8
 
9
9
  ```sh
10
+ node .joycraft/runtime/completion.mjs submit --queue <queue-path> --spec <id>
10
11
  node .joycraft/runtime/completion.mjs scope --queue <queue-path> --spec <id>
11
12
  node .joycraft/runtime/completion.mjs snapshot --queue <queue-path> --spec <id> --out <subject.json>
12
13
  node .joycraft/runtime/completion.mjs finalize-review --queue <queue-path> --spec <id> --report <report.json>
@@ -15,7 +16,9 @@ node .joycraft/runtime/completion.mjs accept-waiver --queue <queue-path> --spec
15
16
  node .joycraft/runtime/completion.mjs recover --queue <queue-path>
16
17
  ```
17
18
 
18
- `scope` checks a queue's declared change scope without writing a subject. `snapshot` does the same check first, then exclusively creates a new `--out` subject file without changing a status. It binds the implementation revision, queue/spec identity, criteria, and SHA-256 values for exact relevant inputs. The shared record shapes are defined in [.joycraft/contracts/types.d.ts](../contracts/types.d.ts). New subjects carry `binding.version: 1`: their queue projection contains the reviewed spec and its transitive dependencies, with completion metadata (`status`, `acceptedReport`, `ownerAcceptance`, and `waiver`) removed from those records. An unrelated queue record can change without staling that subject; a selected spec's other metadata and dependency edges remain bound. A spec binding preserves raw bytes except for its top-level `status:` value.
19
+ `submit` moves a `todo` spec to `in-review` in the queue and frontmatter with one journaled write, recording `submitted: { at }`. It warns about each criterion with no `[AC:<id>]` test, leaves an `in-review` spec unchanged, and exits 2 on a conflict, a `done` spec or an unfinished dependency.
20
+
21
+ `scope` checks a queue's declared change scope without writing a subject. `snapshot` does the same check first, then exclusively creates a new `--out` subject file without changing a status. It binds the implementation revision, queue/spec identity, criteria, and SHA-256 values for exact relevant inputs. The shared record shapes are defined in [.joycraft/contracts/types.d.ts](../contracts/types.d.ts). New subjects carry `binding.version: 1`: their queue projection contains the reviewed spec and its transitive dependencies, with completion metadata (`status`, `submitted`, `acceptedReport`, `ownerAcceptance`, `waiver`) removed from those records. An unrelated queue record can change without staling that subject; a selected spec's other metadata and dependency edges remain bound. A spec binding preserves raw bytes except for its top-level `status:` value.
19
22
 
20
23
  An optional `verification.changeScope` records an explicit full Git commit SHA in `baseRevision`; a branch name, `origin/main`, or inferred base is rejected. Changed verification inputs are captured from NUL-delimited Git status records, including staged, unstaged, rename, and deletion topology. Any changed path outside the verification inputs must be named exactly in `exclusions` with its reason. Store temporary packets and subjects in the project's ignored local-artifact home when it has one; otherwise add each exact untracked snapshot packet output to `exclusions`. `complete` filters only its exact report, receipt, owner-decision, and pending-journal paths while rechecking scope. A deliberate deletion needs a non-requirement verification input with `state: "absent"`; requirement inputs cannot be absent. A rename needs the new present input and the old non-requirement input declared absent.
21
24
 
@@ -27,13 +30,15 @@ Use the original `implementationRevision` as provenance on a re-check while reco
27
30
 
28
31
  ## Independent report and guarded acceptance
29
32
 
30
- The report must use the shared schema and bind its subject identity. It records the actual implementer/reviewer native identities, `independent: true` only when a separate reviewer context actually reviewed the work, scope review, flow evidence, receipts, limits, unresolved items, and a PASS/FAIL/UNVERIFIED verdict.
33
+ The report must use the shared schema and bind its subject identity. It records the actual implementer/reviewer native identities, `independent: true` only when a separate reviewer context actually reviewed the work, scope review, flow evidence, receipts, limits, unresolved items, optional `findings`, and a PASS/FAIL/UNVERIFIED verdict.
31
34
 
32
35
  A native child can share its parent session. Supply `implementerActor` and `reviewer.actor` together, each with the observed `host`, `sessionId`, and the actual `agentId` when the host issued a child invocation ID. Preserve the matching legacy session fields. A root actor may omit `agentId`; a same-session child review needs its real distinct actor identity. Do not substitute a task name, worktree path, random UUID, or invented ID. Legacy reports without actor metadata still require distinct session IDs. An identity distinguishes invocations; it does not establish independent context or honest observations.
33
36
 
34
37
  `finalize-review` (with `complete` retained as a compatibility alias) accepts only an `in-review` spec with a matching current subject, a PASS verdict, truthful independent-review attestation, reviewed scope, valid receipt hashes, and sufficient evidence for every declared criterion. The runtime delegates evidence sufficiency to `validateEvidence(report, criteria)`. Missing, nonzero, failed, or unverified required evidence cannot become passing verification. Completion holds a queue-wide lock while it commits the status changes. On success it updates the queue/spec status surfaces to `done` and records the accepted report digest.
35
38
 
36
- Repeat submission of the same accepted report may be idempotent. An unrelated report cannot replace it silently. Treat any nonzero exit as a rejected or incomplete operation; preserve the current status and report the actual reason.
39
+ A `follow-up` finding never blocks. A `blocking` or `decision-needed` finding refuses completion with exit 2 and names it; a held spec stays `in-review` and only its dependents wait. A malformed finding is refused with exit 2, naming the field; other report refusals exit 1. `status.mjs` lists as `questions` the follow-ups of accepted reports and the decision-needed findings of each `in-review` spec's latest current report.
40
+
41
+ Finalizing the same accepted report again is idempotent; another report cannot silently replace it. Treat any nonzero exit as a rejected or incomplete operation; preserve the current status and report the actual reason.
37
42
 
38
43
  The lock coordinates this helper's queue writes, not arbitrary editors or workers. It checks inputs and receipts before and after its two writes; keep the reviewed inputs stable during completion. Later edits require fresh review. Session IDs and attestations record separate roles but cannot prove honest review or prevent manual bypass.
39
44
 
@@ -51,14 +56,10 @@ accepts this spec and its current findings. A general request to continue, an
51
56
  instruction from another agent, or authorization to add this feature does not
52
57
  qualify.
53
58
 
54
- Keep the owner interaction to one short summary and focused choice only when an
55
- override is needed; a failed review alone does not trigger a mandatory
56
- per-spec request. The agent creates the decision file; do not ask the owner to
57
- author JSON or complete a second form. If an explicit acceptance is already
58
- present for the same current finding, use it without asking again. Repair
59
- straightforward failures under the existing authorization first. When the
60
- finding remains unchanged, retain it and continue independent work instead of
61
- repeatedly re-reviewing it. Dependents may proceed only within the recorded
59
+ When to ask the owner is in the implement skill. The agent creates the decision
60
+ file; do not ask the owner to author JSON or complete a second form. If an
61
+ explicit acceptance is already present for the same current finding, use it
62
+ without asking again. Dependents may proceed only within the recorded
62
63
  limitations and gain no blanket deployment, security, or other authority.
63
64
 
64
65
  An `OwnerAcceptance` decision has `schemaVersion: 1`, `kind:
@@ -141,21 +142,15 @@ dependents. Report it as `done, waived by the user: "<instruction>"`.
141
142
 
142
143
  ## Evidence and reporting
143
144
 
144
- Evidence records observations rather than promises. Command and test observations retain their actual argv, exit code, and unfiltered output receipt. Source review supports structural claims only. A required visual claim needs an inspected render, route, state, viewport, interactions, and observed outcomes. Record a check that should have run and did not (a tool break, an app that did not start) as `UNVERIFIED`, failures as `FAIL`, and limits explicitly. A claim that needs evidence beyond the recorded tool's `canShow`, or a package the user skipped, is an agreed limit, not `UNVERIFIED`; a check within `canShow` that could not run because the tool broke is `UNVERIFIED` with `cause: 'tool-break'`.
145
+ Evidence records observations rather than promises: a command's actual argv, exit code and unfiltered output receipt; structural claims only from source review; an inspected render with route, state, viewport, interactions and outcomes for a visual claim. A check that should have run and did not is `UNVERIFIED` (with `cause: 'tool-break'` when the tool broke), a failure is `FAIL`, and a claim beyond the recorded tool's `canShow`, or in a package the user skipped, is an agreed limit.
145
146
 
146
147
  Report the command result, reviewer identity, receipt paths, actual observations, and limitations. A green suite or an implementation commit alone is not accepted completion.
147
148
 
148
149
  ## Interrupted writes and recovery
149
150
 
150
- The helper uses `.joycraft-completion.pending.json` in the queue directory as a bounded local journal for the queue status, spec frontmatter status, and accepted-report receipt. If it detects an interrupted or conflicting write, it stops and prints recovery information. A normal caught failure rolls back when safe but retains the journal for inspection. Do not hand-edit the files to force completion.
151
-
152
- Inspect the journal and affected project-relative files first. Then run:
153
-
154
- ```sh
155
- node .joycraft/runtime/completion.mjs recover --queue <queue-path>
156
- ```
151
+ Each status write is journaled in `.joycraft-completion.pending.json` beside the queue. A caught failure rolls back when safe and keeps the journal unless nothing was written; an interrupted one leaves it. Do not hand-edit the files to force completion.
157
152
 
158
- Recovery preflights both captured files before changing either: each current file must match the journal's before or after hash. It restores only the recorded before bytes, leaves conflicts untouched, and removes the journal only after a successful restore. Replaying recovery is safe until that journal is removed. Preserve the journal and rejected report for diagnosis. Retry `complete` or `accept-owner` only while its original review and any owner instruction remain current; otherwise obtain fresh evidence and authorization appropriate to the changed scope. If recovery reports a conflict or invalid journal, stop and report the local recovery blocker.
153
+ Inspect the journal and affected files, then run `recover`. It restores the before bytes only when each file matches the journal's before or after hash, and removes the journal only after that, so replaying it is safe. Preserve the journal and rejected report for diagnosis. Retry `complete` or `accept-owner` only while its original review and any owner instruction remain current; otherwise obtain fresh evidence and authorization appropriate to the changed scope. If recovery reports a conflict or invalid journal, stop and report it.
159
154
 
160
155
  Host names compare without letter-case differences; session and agent IDs stay
161
156
  opaque and case-exact. IDs must be nonempty, bounded, trim-stable and free of
@@ -43,19 +43,17 @@ and a snippet of at most 1,800 UTF-8 bytes around the first matching line.
43
43
  Snippets include source line numbers and whether they are partial. Large files
44
44
  are never described as searched or ranked. Read those separately when needed.
45
45
 
46
- Every selected candidate is retained, including a zero-scored file. Jev scores
47
- all selected candidates in batches of at most 20; there is no candidate cap or
48
- score cutoff. Encoded states over 60,000 bytes split into smaller batches, leaving
49
- room for the shared question rubric. Narrow the query or paths before calling when a large area would
50
- create unnecessary calls. A keyword search can miss paraphrases or dependencies;
51
- this result never claims exhaustive context. Refine and follow relevant links.
52
-
53
- The JSON `references` are sorted by Jev relevance, then local score/path. Each
54
- reference identifies `provenance`, `relevance` (or `null`), and its scoring batch.
55
- `batches` records candidate IDs, returned model, usage and elapsed time for each
56
- successful call. `omitted`, `unreadable`, and `unranked` explain files that were
57
- not scored. Hashes bind the result to the bytes read, not future edits. Open the
46
+ Every selected candidate is retained, including a zero-scored file; there is no
47
+ candidate cap or score cutoff. A keyword search can miss paraphrases or
48
+ dependencies; this result never claims exhaustive context. Refine and follow
49
+ relevant links.
50
+
51
+ The JSON `references` come in local-search order: local score, then path.
52
+ `status` is `local-search-only`, and `unranked` lists each candidate with the
53
+ reason `scoring-disabled`. `omitted` and `unreadable` explain files that were
54
+ not read. Hashes bind the result to the bytes read, not future edits. Open the
58
55
  current source before relying on it; a bounded snippet is not the full file.
56
+ Invalid inputs or an unavailable Git inventory exit nonzero.
59
57
 
60
58
  For an unclear state, data, or responsibility boundary, use the result as a
61
59
  starting point and inspect the actual coupled sources as well: the owner and its
@@ -63,28 +61,7 @@ callers, readers, writers, or persistence boundary as applicable. Add discovered
63
61
  paths to the brief/spec verification inputs before dispatch, and state any
64
62
  unresolved coverage. Read `architecture.md` for focused examples. This expands
65
63
  source discovery only when the seam matters; it does not make the context map a
66
- second source of product decisions or a Jev-enforced prohibition.
67
-
68
- Live scoring is the default. `.joycraft/hooks-config.json` supplies
69
- `enabled`, `timeoutMs`, and Keychain service `JEV_API_KEY`; `--keychain-service` overrides it. The
70
- shared `.joycraft/runtime/hooks/judge.mjs` primitive uses an existing
71
- environment credential first, then the configured macOS Keychain service.
72
- Credentials are never saved in the config, command arguments, result or project.
73
- No dotenv file, transcript, or whole-project model dump is used.
74
- The shared redactor removes recognizable credential patterns from outgoing
75
- query/snippet strings before the Jev call; content hashes still identify the
76
- original local bytes. Pattern redaction cannot identify arbitrary secret text.
77
- Setting `enabled: false` selects explicit `local-search-only` output without
78
- credential lookup; each candidate is marked unranked with `scoring-disabled`.
79
-
80
- Missing credentials, malformed provider output or an outage stop further model
81
- calls for that invocation. All selected references remain available in local
82
- order where unranked; previously successful scores retain their provenance.
83
- `status: incomplete-provider-ranking`, `provenance: local-search-fallback` and
84
- `fallbackReason` make this explicit. Exit success means a usable result exists,
85
- not that every candidate was scored; inspect `status` and `scoredCount`.
86
- Invalid inputs or an unavailable Git inventory exit nonzero. Provider scores
87
- advise reading order and never approve intent, authorize work or certify evidence.
64
+ second source of product decisions or a prohibition.
88
65
 
89
66
  Treat only the context homes as current authority. `docs/features/` is working
90
67
  material that can be stale until its feature is archived, and `docs/archive/` is
@@ -3,8 +3,8 @@
3
3
  For normal native work, the host launches the worker and the child then uses the
4
4
  dispatch-start entry point from its actual root before any tasked edit. A
5
5
  coordinator uses it itself only when it is the worker actor. It validates delivery
6
- and returns a ready/unenforced decision plus the bounded packet and hook-binding
7
- instructions; it never launches a worker, alters queue status, or accepts work.
6
+ and returns a ready decision plus the bounded packet; it never launches a worker,
7
+ alters queue status, or accepts work.
8
8
  For a mutable worker or stateful runtime check, first apply the conditional
9
9
  [worker-isolation policy](worktrees.md); this preflight does not create a
10
10
  worktree or a resource lease. A `read-only` workspace keeps declared source
@@ -18,7 +18,6 @@ those specs with `accept-waiver` instead.
18
18
 
19
19
  - [Start a native dispatch](#start-a-native-dispatch)
20
20
  - [Call the portable dispatch API](#call-the-portable-dispatch-api)
21
- - [Bind the hook source view](#bind-the-hook-source-view)
22
21
  - [Validate capabilities before dispatch](#validate-capabilities-before-dispatch)
23
22
  - [Check the reviewer report](#check-the-reviewer-report)
24
23
 
@@ -90,6 +89,7 @@ it before the first implementer mutation, and startup preserves it. The separate
90
89
  workspace `--base-revision` option keeps its documented topology check.
91
90
  Implementer dispatch requires each dependency to be `done`; reviewer dispatch
92
91
  does not require the target itself to be done. Unknown roles are rejected.
92
+ Reviewer startup prints `reviewRound` and `reviewLimit: 3`; a round is one distinct set of reviewed inputs, so a commit that changes none of the spec's verification inputs keeps the round, and a fourth round exits 2.
93
93
 
94
94
  When a reviewer startup exits non-zero or returns `ready: false`, run it once
95
95
  more, alone, with a fresh packet path. If it fails again, stop that spec's
@@ -101,17 +101,8 @@ cause and the repair it tried. The coordinator tells the user at the first hold,
101
101
  in the conversation, while other workers keep running; it does not wait for the
102
102
  session end. The session-end summary lists every held spec again.
103
103
 
104
- The return is `ready: true` for an available valid startup path, including a
105
- hooks-disabled/manual path. It is not a cryptographic attestation. Missing human
106
- prompt history for an enforced startup is explicit
107
- `ready: false`, exit code 2, and `unenforced`; it is not a permission request,
108
- an invented prompt, or a substitute session ID. Continue discovery and other
109
- unrelated authorized work, including manual/disabled workflows, but do not
110
- dispatch the claimed enforced worker. Do not
111
- select an arbitrary task to make a native relationship appear valid. A child
112
- binding uses its own root and a real prompt; it may inherit the same native
113
- parent session authority only when an observed linked coordinator relationship
114
- is valid. A unique new root session needs its own actual prompt.
104
+ A valid startup returns `hooks: { enabled: false, reason: "hooks-not-installed" }`
105
+ and `ready: true`. It is not a cryptographic attestation.
115
106
 
116
107
  When the host reports a native agent/invocation ID, also pass
117
108
  `--implementer-host <host> --implementer-agent <actual-agent-id>`. For a root
@@ -142,8 +133,8 @@ from a legacy self-attested call.
142
133
 
143
134
  Call the portable API from a Codex, Claude, Pi, OMP, or Copilot adapter. Normal
144
135
  native startup calls `startDispatch(options, nativeEvent)` so the runtime can
145
- validate the observed session, agent relationship, and prompt authority before
146
- it returns a packet:
136
+ validate the observed session and agent relationship before it returns a
137
+ packet:
147
138
 
148
139
  ```js
149
140
  import { startDispatch } from './.joycraft/runtime/dispatch-start.mjs';
@@ -171,30 +162,6 @@ When `AGENTS.md` has a `## Review loop` section or this machine has
171
162
  `{ path, section, local? }`, the shared section plus, when present,
172
163
  `local: { path, record }` with this machine's host tools.
173
164
 
174
- ## Bind the hook source view
175
-
176
- Dispatch-start prepares and binds its hash-bound `hookSourceView` with actual
177
- native IDs. The lower-level manual bind command remains only for an explicit
178
- non-native boundary. It uses the actual native session ID reported at startup,
179
- never a generated UUID:
180
-
181
- ```sh
182
- node .joycraft/runtime/hooks/cli.mjs bind-source \
183
- --session <actual-native-session-id> --packet <project-relative-packet.json>
184
- ```
185
-
186
- Bind a known parent session before launching children when the host exposes
187
- parent inheritance. Otherwise, bind the child after it reports its own actual
188
- session ID. For children sharing a parent session but working on different
189
- specs, add `--agent <actual-native-child-id>` when binding each child. Its task
190
- view remains separate while human prompts still inherit from the parent.
191
- Rebind after the brief changes. The hook view currently receives only
192
- hash-checked brief and decision pointers; the worker packet retains the exact
193
- decision records. An unknown ID leaves the dependent
194
- Jev check advisory and unverified; it does not permit a substitute ID. The view
195
- makes selected saved answers available to Jev, but never grants permission or
196
- an expected review outcome.
197
-
198
165
  ## Validate capabilities before dispatch
199
166
 
200
167
  `claims.capabilities` must include `filesystem:read`,
@@ -0,0 +1,60 @@
1
+ # Gates
2
+
3
+ Joycraft's rule is that only an accepted review completes a spec. The gate
4
+ check makes the rule a command, so it does not depend on the agent.
5
+
6
+ ## What the gate checks
7
+
8
+ `node .joycraft/runtime/gate.mjs check` exits 2 and names each problem on
9
+ stderr. Each message names its next step:
10
+
11
+ - a `done` spec with no accepted report, owner acceptance or waiver: set it
12
+ back to `in-review`, get an independent review, then run the completing
13
+ command that the message names;
14
+ - a spec whose queue status and frontmatter status are different: pending
15
+ journal: run the named `recover` command, stop on a conflict; accepted: run
16
+ `node .joycraft/runtime/status.mjs --fix`; unaccepted `done`: the step above
17
+ first; else set the frontmatter to the queue's status and tell the user;
18
+ - a current queue or spec file that does not parse: repair it and tell the user.
19
+
20
+ A spec held at `in-review` is never a problem. Queues older than
21
+ `schemaVersion: 1` are skipped. With no problem, it exits 0 and prints nothing.
22
+ Any agent can run the check, and the session-end skill runs it. Without a hook,
23
+ that step depends on the agent.
24
+
25
+ ## The Stop hook
26
+
27
+ The hook is for Claude Code only for now. When it is on, Claude Code runs the
28
+ gate check each time it is about to stop. On exit 2, Claude Code gives the
29
+ message to the agent, and the agent must do the step that the message names.
30
+ The hook is not a sandbox: a shell command can still write the queue file. The
31
+ check finds that result at the next stop.
32
+
33
+ ## Turn the gates on
34
+
35
+ The user says "Turn on Joycraft's gates", or says yes to the offer in setup.
36
+
37
+ 1. Run `node .joycraft/runtime/gate.mjs hook --plan`. It writes nothing.
38
+ - `change: "none"` and a `message`: tell the user the message. On a host
39
+ other than Claude Code, the message says that the hook is not available
40
+ and how to run the check by hand.
41
+ - `change: "add"`: continue.
42
+ 2. Show the change to `.claude/settings.json` (`before` and `after`) on a page,
43
+ as [showing a page](artifacts.md#showing-a-page) says. In two lines, say
44
+ what the Stop hook does and how to turn it off.
45
+ 3. On the user's yes, run `node .joycraft/runtime/gate.mjs hook --on`. Do not
46
+ edit `.claude/settings.json` by hand.
47
+
48
+ `.claude/settings.json` is the user's file. `--on` adds one Stop hook whose
49
+ command runs `gate.mjs check`, and it changes no other byte. A second `--on`
50
+ changes nothing. If the file is not valid JSON, the command stops and writes
51
+ nothing; tell the user.
52
+
53
+ ## Turn the gates off
54
+
55
+ The user says "Turn off Joycraft's gates". That request is the yes. Run
56
+ `node .joycraft/runtime/gate.mjs hook --off` and tell the user its message.
57
+ When nobody changed the file after `--on`, the file is again byte for byte as
58
+ it was before `--on`. `.joycraft/local/gate-hook.json` keeps the record for
59
+ this. Without a usable record, it removes Joycraft's Stop hook and any
60
+ `hooks.Stop` list or `hooks` object left empty, and its message names each one.
@@ -2,7 +2,7 @@
2
2
 
3
3
  Use `joycraft-retro` on a requested session worth learning from, including a
4
4
  successful session with unexpected friction. It proposes environment improvements;
5
- it does not run on every session end, change acceptance or add Jev calls.
5
+ it does not run on every session end or change acceptance.
6
6
 
7
7
  ## Prepare native session evidence
8
8
 
@@ -22,9 +22,9 @@ node .joycraft/runtime/retro.mjs prepare --source /absolute/path/session.jsonl -
22
22
  ```
23
23
 
24
24
  Line selection reduces the prepared context, not source-read memory: this small
25
- helper reads the supplied file into memory. Use an existing episode export or
26
- the evidence collector for sources too large for that approach, retaining the
27
- original source references and any selection limits.
25
+ helper reads the supplied file into memory. Use an existing episode export for
26
+ sources too large for that approach, retaining the original source references
27
+ and any selection limits.
28
28
 
29
29
  Read `coverage.json` first, then `conversation.md` and `tools.jsonl`. The source
30
30
  hash and physical line references let you return to exact raw evidence. The
@@ -45,17 +45,8 @@ inspect material before sharing it. It creates no permission or decision state.
45
45
  Other harness formats can be read directly; do not describe an unsupported
46
46
  format as normalized or invent native coverage.
47
47
 
48
- Keep any linked session/worker files, hook receipts and final review reports as
49
- separate sources. A single file may not contain all workers or continuations.
50
- Existing hook metrics can summarize explicit receipts without replaying paid
51
- checks:
52
-
53
- ```sh
54
- node .joycraft/runtime/hooks/metrics.mjs --log /absolute/path/events.jsonl
55
- ```
56
-
57
- Provider-reported usage is not an invoice; absent calls, timing or usage remain
58
- unknown. Do not equate a missing receipt with a free or absent call.
48
+ Keep any linked session/worker files and final review reports as separate
49
+ sources. A single file may not contain all workers or continuations.
59
50
 
60
51
  ## Judge a candidate
61
52
 
@@ -49,9 +49,12 @@ On "uninstall Joycraft" or "remove Joycraft":
49
49
  1. Read `.joycraft/installation.json`. Its `files` list holds each installed path
50
50
  and the `sha256` it had at install.
51
51
  2. Show the user what goes: the listed files, with each file whose bytes changed
52
- since install marked as edited, and `.joycraft/` itself. Say what stays:
53
- `AGENTS.md`, `CLAUDE.md`, `docs/` and `joycraft.yaml` are theirs.
52
+ since install marked as edited, Joycraft's Stop hook in `.claude/settings.json`
53
+ when the gates are on, and `.joycraft/` itself. Say what stays: `AGENTS.md`,
54
+ `CLAUDE.md`, `docs/` and `joycraft.yaml` are theirs.
54
55
  3. Offer to remove the Joycraft lines from `AGENTS.md` and `CLAUDE.md` (the context
55
56
  map rows that point into `.joycraft/`, and the Joycraft rules). Show that diff
56
57
  separately.
57
- 4. On the user's yes, delete the files, then `.joycraft/`. Do not commit.
58
+ 4. On the user's yes, first run `node .joycraft/runtime/gate.mjs hook --off`. It
59
+ removes Joycraft's Stop hook. When the gates are off, it changes nothing. Then
60
+ delete the files, then `.joycraft/`. Do not commit.
@@ -320,15 +320,15 @@ the owning brief and do not ask it again.
320
320
  ## Finish setup and continue
321
321
 
322
322
  Step 5 of [Set up the context map](#set-up-the-context-map) is the finish
323
- check; also check that command claims match their cited source. Setup does not approve hooks, change credentials, install
324
- dependencies, or replace the user's workflow. Read `hooks-install.md` only when
325
- hook activation is part of the requested setup.
323
+ check; also check that command claims match their cited source. Setup does not
324
+ change credentials, install dependencies, or replace the user's workflow.
325
+ When `.joycraft/installation.json` lists the `claude` host, offer the [gates](gates.md) once, in one line; turn them on only on the user's yes.
326
326
 
327
327
  ## Upgrade and agent onboarding
328
328
 
329
329
  Updates go through the Joycraft CLI; follow [update guidance](update.md). The
330
330
  installation manifest `.joycraft/installation.json` records the installed bytes;
331
- it is dispatch input, not proof that hooks ran or that a review was independent.
331
+ it is dispatch input, not proof that a review was independent.
332
332
 
333
333
  A new worker reads `.joycraft/preferences.json` before work, reuses explicit
334
334
  model, effort, role, concurrency, verification and question-style choices, and inspects the
@@ -21,30 +21,30 @@ end-to-end requirement follows from these examples.
21
21
 
22
22
  ### Running tests
23
23
 
24
- Run focused checks once, in the runner's non-interactive single-run mode (for
25
- example `vitest run`, `jest --watchAll=false`, or the project script with that
26
- flag passed through) with `CI=1` set. Never start watch mode. Give each test
27
- command a 300000 ms tool timeout or the project's available supervisor. A
28
- timeout while other test runs share the machine is not yet a hang: rerun once,
29
- alone, with fewer workers. If it times out again, stop leftover test processes
30
- and report the condition with the command; do not keep rerunning. Never wait
31
- in a foreground loop (`until …; do sleep`, `kill -0`, polling a log): start a
32
- long command in the background, keep working, and act on its completion
33
- notice. If a shared lock or verify slot stays busy for more than 5 minutes,
34
- stop and report it instead of waiting. Never give a verdict while your own
35
- command is still running. A
36
- reviewer may reuse a suite receipt bound to the same revision instead of
37
- rerunning it. A bug fix needs a meaningful failing regression only when one
38
- can distinguish the repaired fault. Ordinary edits do not require ritual red/green.
24
+ Run each test command once, in single-run mode with `CI=1` and a 300000 ms
25
+ timeout, through
26
+ `node .joycraft/runtime/evidence.mjs run --receipt <path> -- <command...>`
27
+ (`--timeout-ms <n>` changes the limit). It uses no shell and saves raw stdout
28
+ and stderr to a new git-ignored or feature-folder receipt. Never start watch
29
+ mode. Copy the printed `command` and `receipts` into the observation; exit 0
30
+ only means the run was recorded. At the limit it ends the process group:
31
+ `exitCode: null`, `timedOut: true`. Set the host tool timeout to at least this
32
+ limit. A shorter default (for example 120000 ms) ends the wrapper before the
33
+ child. A timeout during parallel runs is not yet a hang: rerun once, alone,
34
+ with fewer workers. If it times out again, report it
35
+ with the receipt; do not keep rerunning. Never wait in a foreground loop
36
+ (`until …; do sleep`, `kill -0`, polling a log): start a long command in the
37
+ background, keep working, and act on its completion notice. If a shared lock
38
+ or verify slot stays busy over 5 minutes, stop and report it. Never
39
+ give a verdict while your own command is still running. A reviewer may reuse
40
+ a suite receipt bound to the same revision. A bug fix needs a meaningful
41
+ failing regression only when one can distinguish the repaired fault.
39
42
 
40
43
  ## Commands and receipt records
41
44
 
42
- For `test` and `command` observations, `command.argv` records the executed command
43
- and `command.exitCode` records its actual outcome. `null` means unavailable.
44
- Keep raw stdout/stderr, including failure output, in a receipt; do not substitute
45
- grepped success text or a pipeline's final status for the command outcome. A
46
- nonzero or missing status cannot support PASS. Do not claim a process finished
47
- while it is still running.
45
+ In `test` and `command` observations, `command.exitCode` is the actual outcome;
46
+ `null` means unavailable. Never substitute grepped success text or a
47
+ pipeline's final status for it. A nonzero or missing status cannot support PASS.
48
48
 
49
49
  Each observation has at least one supporting receipt with a project-relative path
50
50
  and SHA-256. A receipt may be raw command output, inspection notes, an image, or a
@@ -128,7 +128,7 @@ guidance](https://code.claude.com/docs/en/hooks#reference-scripts-by-path).
128
128
 
129
129
  On hosts without that native capability, create an equivalent manual Git linked
130
130
  worktree at an explicit, verified ref. Build each packet from its actual root:
131
- `workerRoot`, manifest path, source paths, hook binding context, commands,
131
+ `workerRoot`, manifest path, source paths, commands,
132
132
  receipts, and report paths must resolve there. A coordinator may not substitute
133
133
  the shared checkout merely because its sources look equivalent.
134
134